Reproduce “You Are What You Say: Exploiting Linguistic Content for VoicePrivacy Attacks” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Ünal Ege Gaznepoglu, Anna Leschanowsky, Ahmad Aloradi, Prachi Singh, Daniel Tenbrinck, Emanuël A. P. Habets, Nils Peters
Researchers from AudioLabs Erlangen and Fraunhofer IIS demonstrate that linguistic content itself, even with acoustic features anonymized, can be exploited for speaker re-identification. They present a BERT-based text-only attack achieving mean Equal Error Rates of 33-36%, comparable to acoustic attacks, and as low as 1.6% for specific individuals, with explainability analysis attributing this to consistent semantic keywords.
Key insights
- Global EERs in speaker anonymization benchmarks can be misleading, as they mask significant speaker-specific vulnerabilities, advocating for more granular, speaker-level privacy evaluations.
- Intra-speaker linguistic content similarity, such as recurring semantic themes or keywords, can be exploited to re-identify speakers even when acoustic features are anonymized.
- The widely used LibriSpeech dataset, curated from audiobooks, inherently contains linguistic biases (e.g., speakers reading from specific topics) that make certain individuals identifiable through their textual content.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.